Waarom GDPR niet deugt

Nog een fragment uit het interview in Der Spiegel met AI-expert Pedro Domingos. De nieuwe Europese regelgeving die onze privacy zou moeten beschermen, de General Data Protection Regulation (GDPR), werkt niet omdat hij zich te veel richt op begrijpelijkheid. Algoritmes moeten voor iedereen te begrijpen zijn anders zouden ze kunnen worden misbruikt. Maar dat is onmogelijk, zegt Domingos, en ook niet wenselijk. Veel AI kan nu al dingen die mensen niet begrijpen, en dat is niet erg; veel wetenschappelijke ontdekkingen zijn ook per ongeluk gedaan, in onwetendheid (penicilline, röntgenstraling). Beter om te richten op effecten: als Facebook onze data verzamelt om er geld mee te verdienen is dat niet erg – maar je zou Facebook moeten kunnen verplichten om iets tegenover hun gewin te zetten. Bijvoorbeeld: dat ze helpen om het waarheidsgehalte te bepalen van de mededelingen op hun netwerk. (Engelse uitgebreide versie hieronder) 

 Domingos: The European Union’s General Data Protection Regulation (GDPR) is putting too much value on the factor of explainability — meaning why an algorithm decides this way rather than that way. Let’s take the example of cancer research, where machine learning already plays an important role. Would I rather be diagnosed by a system that is 90 percent accurate but doesn’t explain anything, or a system that is 80 percent accurate and explains things? I’d rather go for the 90 percent accurate system.

DER SPIEGEL: Why can’t we have both — accuracy and explainability?

Domingos: The best learning algorithms are these neural network-based ones inspired by what we find in humans and animals. These algorithms are very accurate as they can understand the world based on a lot of data at a much more complex level than we can. But they are completely opaque. Even we, the experts, don’t understand exactly how they work. We only know that they do. So, we should not allow only algorithms which are fully explainable. It is hard to capture the whole complexity of reality and keep things at the same time accurate and simple.

DER SPIEGEL: How, then, should AI be regulated?

Domingos: There is no law detailed enough to compete with the complexity of things that algorithms can do. What can be regulated, though, are the dangers that come from overly crude objective functions such as Facebook’s algorithms maximizing the time you spend on their site. These can be regulated by saying: OK, you have these business elements in your objective function because you need to make money. But you should also have these societal goals like, for example, the truth value of the things that are being said.